First
E. coli
Proof of concept for the artificial replicon, plus high-copy plasmid production strains as a first product. Bacterial plasmids do not need a centromere, so this stage proves copy number and the screening engine.
Technology
Precision Plasmids are built from first principles rather than borrowed from nature. We screen for the genetic parts that make a construct replicate and segregate in a given organism, then assemble them into a vehicle whose copy number and inheritance or persistence are design inputs.
Anatomy
A plasmid that replicates but cannot segregate drifts and is lost. One that segregates but cannot replicate runs out of copies. Precision Plasmids carry engineered versions of both functions, so the construct maintains itself in a dividing cell population without entering the genome. In non-dividing cells, segregation is replaced by persistence, the ability to remain stably in the cell.
An origin and the initiator functions that go with it, so the construct replicates to a copy number of our choice, based on a customer’s needs, rather than whatever replication the host cell happens to trigger.
Sequence that nucleates the host’s own mitotic partitioning machinery, so copies are actively distributed to both daughter cells instead of left to chance.
Any size cargo from single genes to large, multi-gene clusters and pathways. Without a capsid, payload size is a design decision rather than a packaging limit.
Promoters, terminators and regulatory parts matched to the host organism and tuned to the product.
Staying outside the genome removes position effects and silencing at an integration site. Precision Plasmids give consistent expression and performance every time they are introduced into a cell. Once the design is in place, cargo gene expression is consistent and predictable.
The platform
The same high-level design and workflow are used on every organism. We cast a wide net across the cellular functions we need, characterize the screening output, and keep what the screen teaches us.
We build a plasmid design space from our technology design parameters, organism-specific genes, diversity libraries and computational prediction, then assemble it into a library we screen in the target cell.
Combinatorial assembly turns the design space into candidate replicons, centromeres and regulatory parts, carried on screening plasmids.
Ultra-high-throughput screens measure what matters: does it replicate, at what copy number, is it inherited, and does it hold over many divisions.
Every screen pairs sequence with measured function. These pairs become a database and a model, and the next design round is informed by the ones that preceded it, chosen rather than sampled.
The founders built and ran this kind of screening cascade at Primordial Genetics and Primrose Bio, for enzyme and organism improvement. What is new is the target.
Validation path
We develop in order of difficulty and evolutionary complexity. Microbial work moves fast and teaches us the design rules. Yeast is the first real test of a synthetic centromere. Mammalian cells are where the value is, and where we venture with the most data and experience under our belts.
First
Proof of concept for the artificial replicon, plus high-copy plasmid production strains as a first product. Bacterial plasmids do not need a centromere, so this stage proves copy number and the screening engine.
Next
The first test of a synthetic centromere in a eukaryote. Natural yeast centromere plasmids hold one or two copies with no way to set the number. A designed centromere makes it a parameter. Natural centromeres do not work in some yeast species. Our artificial centromeres will.
Key decision pointThen
Proof of concept in cultured mammalian cells. This is the milestone that opens pharma manufacturing at scale, and later, with partners, cell and gene therapy and regenerative medicine.
Head to head
Every delivery vehicle currently on the market is strong in some areas and weak somewhere else. The gap we are building into is high delivery performance, tunability in copy number and stability, with simple, scalable manufacturing. Precision Plasmids™ = no trade-offs.
| Type | Payload | Copy number per cell | Stability | Manufacturing complexity | Notes and drawbacks |
|---|---|---|---|---|---|
| Naked DNAPlasmid DNA | 100 kb and above | Low to high, 10 to 1000+ | Transient | Low | A low share of targeted cells receive the DNA. |
| Naked DNAPCR DNA | Low to medium, up to 10 kb | Low to high, 10 to 1000+ | Transient | Low | A low share of targeted cells receive the DNA. |
| VirusesTransient: AAV, adenovirus, HSV | Low to high, 4.5 to 100+ kb | Low to medium, 0.5 to 100+ | Transient | High | Immunogenicity, toxicity and adverse effects. |
| VirusesIntegrating: lentiviruses, retroviruses | Low to medium, 8 to 10 kb | Low, 1 to 5 | Stable | High | Insertional oncogenesis risk, and prone to epigenetic silencing. |
| Autonomously replicating plasmidsStelli Bio Precision Plasmids™ | 100 kb and above | Tunable | Tunable | Low | Potentially the best of all worlds. |
Precision Plasmid values are engineering design specifications, not measured data. Each is validated in turn. Precision Plasmids™ are being developed in bacteria, then yeast, then mammalian cells.
White space
Plot the field on two axes and the pattern is clear. Technologies that persist without integrating are too large to deliver or manufacture. Technologies that are easy to deliver and make do not persist. Viruses are immunogenic and cause adverse effects.
The upper right quadrant is empty today, which is the reason the company exists.
Select a point to read the detail.
A synthetic replicon and a synthetic centromere at plasmid scale. The design goal is chromosome-like maintenance in a construct that can be delivered and manufactured like a plasmid.
Status: in development, validated organism by organism.
Closest analogHuman and mammalian artificial chromosomes solve inheritance properly, with real centromeres and excellent segregation. But at megabase scale they are not practically deliverable and not manufacturable at commercial scale, so they remain research and animal-model tools.
Great biology, wrong form factor.
AdjacentS/MAR systems and the viral oriP/EBNA1 system stay compact and avoid integration, and EBNA1 does tether episomes to mitotic chromosomes for partial partitioning. But retention is incomplete, copy number is not tunable, and EBNA1 brings a viral protein into the product.
Right form factor, incomplete inheritance.
Today’s defaultLentivirus and transposon systems give durable expression because they insert into the genome. That is also the problem: insertion is semi-random, bringing genotoxicity risk and silencing, and copy number becomes something you screen for.
Durability bought with the genome.
Today’s defaultAAV delivers efficiently and engineered capsids improve targeting. But payload is capped by the capsid, pre-existing immunity excludes many patients, re-dosing is largely blocked, and the episome dilutes out in dividing tissue. Manufacturing is expensive and capacity-limited.
Delivery without persistence or control.
BaselinePlasmid and linear DNA are cheap, simple and well tolerated, which is why they are everywhere in manufacturing. They do not replicate or segregate in mammalian cells, so expression is transient and programs compensate with high or repeated doses and physical delivery methods.
Manufacturable, but it fades.
The data layer
Our screens are built to generate something that does not exist in the literature: sequence paired with measured replication efficiency, copy number and mitotic stability across large designed libraries, including the constructs that fail.
Genomic and protein language models propose candidate origins, initiator proteins and centromere-nucleating sequences that a homology search alone would never surface. The library starts enriched rather than naive.
Every screen produces labeled pairs. Those pairs train models that predict function from sequence, so the next design round is chosen rather than sampled. Negatives are captured too, which is the half of the data no publication contains.
One screening architecture means every organism produces data in the same representation. Learning from bacteria and yeast carries into mammalian design, which is exactly where the technical risk concentrates.
Said plainly: AI does not make the biology work. The screens do. What AI changes is the number of design cycles needed to get there, and what happens to the data afterwards. A screen without a model is a hit list that expires. A screen with a model is a design capability that appreciates.
This section describes the data layer we are building, not assets in place today.
The most useful thing a partner can bring us is a real constraint. Organism, payload size, target copy number, generations of stability.